Update Microsoft.Extensions.AI to 9.3.0-preview.1.25114.11 (#7388)

stephentoub · web-flow · commit 8af76acd0b1e · 2025-02-20T11:16:25.000-08:00
diff --git a/docs/samples/Microsoft.ML.GenAI.Samples/MEAI/Llama3_1.cs b/docs/samples/Microsoft.ML.GenAI.Samples/MEAI/Llama3_1.cs
@@ -46,7 +46,7 @@ public static async Task RunAsync(string weightFolder, string checkPointName = "
             """;
         var chatMessage = new ChatMessage(ChatRole.User, task);
 
-        await foreach (var response in client.CompleteStreamingAsync([chatMessage]))
+        await foreach (var response in client.GetStreamingResponseAsync([chatMessage]))
         {
             Console.Write(response.Text);
         }
diff --git a/docs/samples/Microsoft.ML.GenAI.Samples/MEAI/Phi3.cs b/docs/samples/Microsoft.ML.GenAI.Samples/MEAI/Phi3.cs
@@ -36,7 +36,7 @@ public static async Task RunAsync(string weightFolder)
             """;
         var chatMessage = new ChatMessage(ChatRole.User, task);
 
-        await foreach (var response in client.CompleteStreamingAsync([chatMessage]))
+        await foreach (var response in client.GetStreamingResponseAsync([chatMessage]))
         {
             Console.Write(response.Text);
         }
diff --git a/eng/Versions.props b/eng/Versions.props
@@ -46,7 +46,7 @@
     <MicrosoftDotNetInteractiveVersion>1.0.0-beta.24375.2</MicrosoftDotNetInteractiveVersion>
     <MicrosoftMLOnnxRuntimeVersion>1.18.1</MicrosoftMLOnnxRuntimeVersion>
     <MlNetMklDepsVersion>0.0.0.12</MlNetMklDepsVersion>
-    <MicrosoftExtensionsAIVersion>9.0.1-preview.1.24570.5</MicrosoftExtensionsAIVersion>
+    <MicrosoftExtensionsAIVersion>9.3.0-preview.1.25114.11</MicrosoftExtensionsAIVersion>
     <!-- runtime.native.System.Data.SqlClient.sni is not updated by dependency flow as it is not produced live anymore. -->
     <RuntimeNativeSystemDataSqlClientSniVersion>4.4.0</RuntimeNativeSystemDataSqlClientSniVersion>
     <!--
diff --git a/src/Microsoft.ML.GenAI.Core/CausalLMPipelineChatClient.cs b/src/Microsoft.ML.GenAI.Core/CausalLMPipelineChatClient.cs
@@ -20,21 +20,20 @@ public abstract class CausalLMPipelineChatClient<TTokenizer, TCausalLMModel> : I
 {
     private readonly ICausalLMPipeline<TTokenizer, TCausalLMModel> _pipeline;
     private readonly IMEAIChatTemplateBuilder _chatTemplateBuilder;
+    private readonly ChatClientMetadata _metadata;
 
     public CausalLMPipelineChatClient(
         ICausalLMPipeline<TTokenizer, TCausalLMModel> pipeline,
         IMEAIChatTemplateBuilder chatTemplateBuilder,
         ChatClientMetadata? metadata = null)
     {
         var classNameWithType = $"{nameof(CausalLMPipelineChatClient<TTokenizer, TCausalLMModel>)}<{typeof(TTokenizer).Name}, {typeof(TCausalLMModel).Name}>";
-        Metadata ??= new ChatClientMetadata(providerName: classNameWithType, modelId: typeof(TCausalLMModel).Name);
+        _metadata = new ChatClientMetadata(providerName: classNameWithType, modelId: typeof(TCausalLMModel).Name);
         _chatTemplateBuilder = chatTemplateBuilder;
         _pipeline = pipeline;
     }
 
-    public ChatClientMetadata Metadata { get; }
-
-    public virtual Task<ChatCompletion> CompleteAsync(IList<ChatMessage> chatMessages, ChatOptions? options = null, CancellationToken cancellationToken = default)
+    public virtual Task<ChatResponse> GetResponseAsync(IList<ChatMessage> chatMessages, ChatOptions? options = null, CancellationToken cancellationToken = default)
     {
         var prompt = _chatTemplateBuilder.BuildPrompt(chatMessages, options);
         var stopSequences = options?.StopSequences ?? Array.Empty<string>();
@@ -46,15 +45,15 @@ public virtual Task<ChatCompletion> CompleteAsync(IList<ChatMessage> chatMessage
             stopSequences: stopSequences.ToArray()) ?? throw new InvalidOperationException("Failed to generate a reply.");
 
         var chatMessage = new ChatMessage(ChatRole.Assistant, output);
-        return Task.FromResult(new ChatCompletion([chatMessage])
+        return Task.FromResult(new ChatResponse([chatMessage])
         {
             CreatedAt = DateTime.UtcNow,
             FinishReason = ChatFinishReason.Stop,
         });
     }
 
 #pragma warning disable CS1998 // Async method lacks 'await' operators and will run synchronously
-    public virtual async IAsyncEnumerable<StreamingChatCompletionUpdate> CompleteStreamingAsync(
+    public virtual async IAsyncEnumerable<ChatResponseUpdate> GetStreamingResponseAsync(
 #pragma warning restore CS1998 // Async method lacks 'await' operators and will run synchronously
         IList<ChatMessage> chatMessages,
         ChatOptions? options = null,
@@ -69,7 +68,7 @@ public virtual async IAsyncEnumerable<StreamingChatCompletionUpdate> CompleteStr
             temperature: options?.Temperature ?? 0.7f,
             stopSequences: stopSequences.ToArray()))
         {
-            yield return new StreamingChatCompletionUpdate
+            yield return new ChatResponseUpdate
             {
                 Role = ChatRole.Assistant,
                 Text = output,
@@ -83,6 +82,8 @@ public virtual void Dispose()
     }
 
     public virtual object? GetService(Type serviceType, object? serviceKey = null) =>
-        serviceKey is null && serviceType is not null && serviceType.IsAssignableFrom(GetType()) ? this :
+        serviceKey is not null ? null :
+        serviceType == typeof(ChatClientMetadata) ? _metadata :
+        serviceType.IsAssignableFrom(GetType()) ? this :
         null;
 }
diff --git a/src/Microsoft.ML.GenAI.LLaMA/Llama3CausalLMChatClient.cs b/src/Microsoft.ML.GenAI.LLaMA/Llama3CausalLMChatClient.cs
@@ -24,7 +24,7 @@ public Llama3CausalLMChatClient(
     {
     }
 
-    public override Task<ChatCompletion> CompleteAsync(
+    public override Task<ChatResponse> GetResponseAsync(
         IList<ChatMessage> chatMessages,
         ChatOptions? options = null,
         CancellationToken cancellationToken = default)
@@ -40,10 +40,10 @@ public override Task<ChatCompletion> CompleteAsync(
             options.StopSequences = new List<string> { _eotToken };
         }
 
-        return base.CompleteAsync(chatMessages, options, cancellationToken);
+        return base.GetResponseAsync(chatMessages, options, cancellationToken);
     }
 
-    public override IAsyncEnumerable<StreamingChatCompletionUpdate> CompleteStreamingAsync(
+    public override IAsyncEnumerable<ChatResponseUpdate> GetStreamingResponseAsync(
         IList<ChatMessage> chatMessages,
         ChatOptions? options = null,
         CancellationToken cancellationToken = default)
@@ -52,6 +52,6 @@ public override IAsyncEnumerable<StreamingChatCompletionUpdate> CompleteStreamin
         options.StopSequences ??= [];
         options.StopSequences.Add(_eotToken);
 
-        return base.CompleteStreamingAsync(chatMessages, options, cancellationToken);
+        return base.GetStreamingResponseAsync(chatMessages, options, cancellationToken);
     }
 }
diff --git a/src/Microsoft.ML.GenAI.Phi/Phi3/Phi3CausalLMChatClient.cs b/src/Microsoft.ML.GenAI.Phi/Phi3/Phi3CausalLMChatClient.cs
@@ -29,7 +29,7 @@ public Phi3CausalLMChatClient(
     {
     }
 
-    public override Task<ChatCompletion> CompleteAsync(
+    public override Task<ChatResponse> GetResponseAsync(
         IList<ChatMessage> chatMessages,
         ChatOptions? options = null,
         CancellationToken cancellationToken = default)
@@ -45,10 +45,10 @@ public override Task<ChatCompletion> CompleteAsync(
             options.StopSequences = [_eotToken];
         }
 
-        return base.CompleteAsync(chatMessages, options, cancellationToken);
+        return base.GetResponseAsync(chatMessages, options, cancellationToken);
     }
 
-    public override IAsyncEnumerable<StreamingChatCompletionUpdate> CompleteStreamingAsync(
+    public override IAsyncEnumerable<ChatResponseUpdate> GetStreamingResponseAsync(
         IList<ChatMessage> chatMessages,
         ChatOptions? options = null,
         CancellationToken cancellationToken = default)
@@ -57,6 +57,6 @@ public override IAsyncEnumerable<StreamingChatCompletionUpdate> CompleteStreamin
         options.StopSequences ??= [];
         options.StopSequences.Add(_eotToken);
 
-        return base.CompleteStreamingAsync(chatMessages, options, cancellationToken);
+        return base.GetStreamingResponseAsync(chatMessages, options, cancellationToken);
     }
 }

Original file line number	Diff line number	Diff line change
`@@ -46,7 +46,7 @@ public static async Task RunAsync(string weightFolder, string checkPointName = "`
`46`	`46`	`""";`
`47`	`47`	`var chatMessage = new ChatMessage(ChatRole.User, task);`
`48`	`48`
`49`		`- await foreach (var response in client.CompleteStreamingAsync([chatMessage]))`
	`49`	`+ await foreach (var response in client.GetStreamingResponseAsync([chatMessage]))`
`50`	`50`	`{`
`51`	`51`	`Console.Write(response.Text);`
`52`	`52`	`}`
Original file line number	Diff line number	Diff line change
`@@ -36,7 +36,7 @@ public static async Task RunAsync(string weightFolder)`
`36`	`36`	`""";`
`37`	`37`	`var chatMessage = new ChatMessage(ChatRole.User, task);`
`38`	`38`
`39`		`- await foreach (var response in client.CompleteStreamingAsync([chatMessage]))`
	`39`	`+ await foreach (var response in client.GetStreamingResponseAsync([chatMessage]))`
`40`	`40`	`{`
`41`	`41`	`Console.Write(response.Text);`
`42`	`42`	`}`
Original file line number	Diff line number	Diff line change
`@@ -20,21 +20,20 @@ public abstract class CausalLMPipelineChatClient<TTokenizer, TCausalLMModel> : I`
`20`	`20`	`{`
`21`	`21`	`private readonly ICausalLMPipeline<TTokenizer, TCausalLMModel> _pipeline;`
`22`	`22`	`private readonly IMEAIChatTemplateBuilder _chatTemplateBuilder;`
	`23`	`+ private readonly ChatClientMetadata _metadata;`
`23`	`24`
`24`	`25`	`public CausalLMPipelineChatClient(`
`25`	`26`	`ICausalLMPipeline<TTokenizer, TCausalLMModel> pipeline,`
`26`	`27`	`IMEAIChatTemplateBuilder chatTemplateBuilder,`
`27`	`28`	`ChatClientMetadata? metadata = null)`
`28`	`29`	`{`
`29`	`30`	`var classNameWithType = $"{nameof(CausalLMPipelineChatClient<TTokenizer, TCausalLMModel>)}<{typeof(TTokenizer).Name}, {typeof(TCausalLMModel).Name}>";`
`30`		`- Metadata ??= new ChatClientMetadata(providerName: classNameWithType, modelId: typeof(TCausalLMModel).Name);`
	`31`	`+ _metadata = new ChatClientMetadata(providerName: classNameWithType, modelId: typeof(TCausalLMModel).Name);`
`31`	`32`	`_chatTemplateBuilder = chatTemplateBuilder;`
`32`	`33`	`_pipeline = pipeline;`
`33`	`34`	`}`
`34`	`35`
`35`		`- public ChatClientMetadata Metadata { get; }`
`36`		`-`
`37`		`- public virtual Task<ChatCompletion> CompleteAsync(IList<ChatMessage> chatMessages, ChatOptions? options = null, CancellationToken cancellationToken = default)`
	`36`	`+ public virtual Task<ChatResponse> GetResponseAsync(IList<ChatMessage> chatMessages, ChatOptions? options = null, CancellationToken cancellationToken = default)`
`38`	`37`	`{`
`39`	`38`	`var prompt = _chatTemplateBuilder.BuildPrompt(chatMessages, options);`
`40`	`39`	`var stopSequences = options?.StopSequences ?? Array.Empty<string>();`
`@@ -46,15 +45,15 @@ public virtual Task<ChatCompletion> CompleteAsync(IList<ChatMessage> chatMessage`
`46`	`45`	`stopSequences: stopSequences.ToArray()) ?? throw new InvalidOperationException("Failed to generate a reply.");`
`47`	`46`
`48`	`47`	`var chatMessage = new ChatMessage(ChatRole.Assistant, output);`
`49`		`- return Task.FromResult(new ChatCompletion([chatMessage])`
	`48`	`+ return Task.FromResult(new ChatResponse([chatMessage])`
`50`	`49`	`{`
`51`	`50`	`CreatedAt = DateTime.UtcNow,`
`52`	`51`	`FinishReason = ChatFinishReason.Stop,`
`53`	`52`	`});`
`54`	`53`	`}`
`55`	`54`
`56`	`55`	`#pragma warning disable CS1998 // Async method lacks 'await' operators and will run synchronously`
`57`		`- public virtual async IAsyncEnumerable<StreamingChatCompletionUpdate> CompleteStreamingAsync(`
	`56`	`+ public virtual async IAsyncEnumerable<ChatResponseUpdate> GetStreamingResponseAsync(`
`58`	`57`	`#pragma warning restore CS1998 // Async method lacks 'await' operators and will run synchronously`
`59`	`58`	`IList<ChatMessage> chatMessages,`
`60`	`59`	`ChatOptions? options = null,`
`@@ -69,7 +68,7 @@ public virtual async IAsyncEnumerable<StreamingChatCompletionUpdate> CompleteStr`
`69`	`68`	`temperature: options?.Temperature ?? 0.7f,`
`70`	`69`	`stopSequences: stopSequences.ToArray()))`
`71`	`70`	`{`
`72`		`- yield return new StreamingChatCompletionUpdate`
	`71`	`+ yield return new ChatResponseUpdate`
`73`	`72`	`{`
`74`	`73`	`Role = ChatRole.Assistant,`
`75`	`74`	`Text = output,`
`@@ -83,6 +82,8 @@ public virtual void Dispose()`
`83`	`82`	`}`
`84`	`83`
`85`	`84`	`public virtual object? GetService(Type serviceType, object? serviceKey = null) =>`
`86`		`- serviceKey is null && serviceType is not null && serviceType.IsAssignableFrom(GetType()) ? this :`
	`85`	`+ serviceKey is not null ? null :`
	`86`	`+ serviceType == typeof(ChatClientMetadata) ? _metadata :`
	`87`	`+ serviceType.IsAssignableFrom(GetType()) ? this :`
`87`	`88`	`null;`
`88`	`89`	`}`
Original file line number	Diff line number	Diff line change
`@@ -24,7 +24,7 @@ public Llama3CausalLMChatClient(`
`24`	`24`	`{`
`25`	`25`	`}`
`26`	`26`
`27`		`- public override Task<ChatCompletion> CompleteAsync(`
	`27`	`+ public override Task<ChatResponse> GetResponseAsync(`
`28`	`28`	`IList<ChatMessage> chatMessages,`
`29`	`29`	`ChatOptions? options = null,`
`30`	`30`	`CancellationToken cancellationToken = default)`
`@@ -40,10 +40,10 @@ public override Task<ChatCompletion> CompleteAsync(`
`40`	`40`	`options.StopSequences = new List<string> { _eotToken };`
`41`	`41`	`}`
`42`	`42`
`43`		`- return base.CompleteAsync(chatMessages, options, cancellationToken);`
	`43`	`+ return base.GetResponseAsync(chatMessages, options, cancellationToken);`
`44`	`44`	`}`
`45`	`45`
`46`		`- public override IAsyncEnumerable<StreamingChatCompletionUpdate> CompleteStreamingAsync(`
	`46`	`+ public override IAsyncEnumerable<ChatResponseUpdate> GetStreamingResponseAsync(`
`47`	`47`	`IList<ChatMessage> chatMessages,`
`48`	`48`	`ChatOptions? options = null,`
`49`	`49`	`CancellationToken cancellationToken = default)`
`@@ -52,6 +52,6 @@ public override IAsyncEnumerable<StreamingChatCompletionUpdate> CompleteStreamin`
`52`	`52`	`options.StopSequences ??= [];`
`53`	`53`	`options.StopSequences.Add(_eotToken);`
`54`	`54`
`55`		`- return base.CompleteStreamingAsync(chatMessages, options, cancellationToken);`
	`55`	`+ return base.GetStreamingResponseAsync(chatMessages, options, cancellationToken);`
`56`	`56`	`}`
`57`	`57`	`}`
Original file line number	Diff line number	Diff line change
`@@ -29,7 +29,7 @@ public Phi3CausalLMChatClient(`
`29`	`29`	`{`
`30`	`30`	`}`
`31`	`31`
`32`		`- public override Task<ChatCompletion> CompleteAsync(`
	`32`	`+ public override Task<ChatResponse> GetResponseAsync(`
`33`	`33`	`IList<ChatMessage> chatMessages,`
`34`	`34`	`ChatOptions? options = null,`
`35`	`35`	`CancellationToken cancellationToken = default)`
`@@ -45,10 +45,10 @@ public override Task<ChatCompletion> CompleteAsync(`
`45`	`45`	`options.StopSequences = [_eotToken];`
`46`	`46`	`}`
`47`	`47`
`48`		`- return base.CompleteAsync(chatMessages, options, cancellationToken);`
	`48`	`+ return base.GetResponseAsync(chatMessages, options, cancellationToken);`
`49`	`49`	`}`
`50`	`50`
`51`		`- public override IAsyncEnumerable<StreamingChatCompletionUpdate> CompleteStreamingAsync(`
	`51`	`+ public override IAsyncEnumerable<ChatResponseUpdate> GetStreamingResponseAsync(`
`52`	`52`	`IList<ChatMessage> chatMessages,`
`53`	`53`	`ChatOptions? options = null,`
`54`	`54`	`CancellationToken cancellationToken = default)`
`@@ -57,6 +57,6 @@ public override IAsyncEnumerable<StreamingChatCompletionUpdate> CompleteStreamin`
`57`	`57`	`options.StopSequences ??= [];`
`58`	`58`	`options.StopSequences.Add(_eotToken);`
`59`	`59`
`60`		`- return base.CompleteStreamingAsync(chatMessages, options, cancellationToken);`
	`60`	`+ return base.GetStreamingResponseAsync(chatMessages, options, cancellationToken);`
`61`	`61`	`}`
`62`	`62`	`}`